ram-shreyas-naik-sabavat/SidewalkPilot_v1_and_v2

SidewalkPilot Series 1 and 2 Steering Dataset

SidewalkPilot Series 1 and 2 is the finalized camera-to-steering dataset for the baseline and failure/iteration model series. The dataset pairs real field images with steering servo labels in degrees, so a model can learn to map a camera frame to a steering command.

CARLA-assisted. The Series 1/2 models were trained on a blend of these real field images plus CARLA synthetic frames (down-weighted vs real). This repository holds the real labeled images; the CARLA synthetic set is published separately as SidewalkPilot_carla.

Project code and documentation are maintained in the GitHub repo:

Dataset Contents

Trainer source and tests are versioned only in GitHub so there is one canonical code history. The dataset repository contains data and its card, not duplicate executable code.

Extract the image archive before training or evaluation:

tar -xf sidewalkpilot_v1_and_v2_dataset.tar

This creates sidewalkpilot_dataset/ beside the metadata files. The archive excludes macOS metadata and upload-cache files.

Current Size

Label Format

steering_corrections.json is a JSON list. Each entry points to one image and stores the steering target used for training.

Example entry:

{
  "image": "sidewalkpilot_dataset/photo_20260425_145756.jpg",
  "steering": 110.0,
  "repeat": 50,
  "source": "D0425_street_test"
}

Steering Label Meaning

The steering label is a servo angle in degrees.

Steering Distribution

Source Breakdown

Image Sizes

The training pipeline resizes images before inference/training, so mixed capture resolutions are expected.

Basic Loading Example

from pathlib import Path
import json

dataset_root = Path("sidewalkpilot_dataset")
labels = json.loads(Path("steering_corrections.json").read_text())

first = labels[0]
image_name = Path(first["image"]).name
image_path = dataset_root / image_name
steering_degrees = float(first["steering"])

print(image_path, steering_degrees)

Training Use

The labels are intended for the SidewalkPilot steering trainer. Trainer code is maintained in GitHub, not duplicated in this dataset repository. The current training setup uses the image folder plus steering_corrections.json as the correction/label source.

Typical local training flow:

cd code/ai_models_datasets/series_1_and_2
python3 sidewalkpilot_trainer.py \
  --roots sidewalkpilot_dataset \
  --corrections steering_corrections.json \
  --model-version 2.4

Exact training commands may differ depending on whether CARLA data, source weighting, shadow augmentation, or other augmentation settings are being used.

Evaluation Use

The dataset is used to compare SidewalkPilot model checkpoints on the same labeled image set. Common metrics include:

Intended Scope

This dataset supports the closed Series 1.x and 2.x research cycle. Series 1.x was the baseline working series, while Series 2.x pushed the same steering-only architecture to its limits and recorded the failure/iteration data used to design Series 3.x.

This dataset is finalized for the Series 1/2 Hugging Face release. New steering+throttle data should go into the separate Series 3 dataset instead.